given by combine_tfa_policies_lib.get_input_signature()
and action spec given by combine_tfa_policies_lib.get_action_spec()
The combiner policy uses a new timestep spec feature "model_selector"
to select the requested policy at the current state. The feature is
computed as a md5 hash from the respective policies names.

…iler-opt into policy_combiner

TimeStep. The patch also gives the option to keep the temporary
working_dir by setting keep_temps in compilation_runner.py to a
directory where all temporary working_dirs will be saved.

compiles only using iclang instead of both clang
and iclang.

Changed the unit test for MLGOEnvironment to check if
the clang sessions are equal and respect the interactive_only
variable.

…piler-opt into behavior_cloning